TenderGPT Edge Is Evolving: Introducing TenderLM
TenderGPT Edge started as an experiment to answer a simple question:
Can useful AI-powered tender intelligence run locally on an ordinary laptop?
Since then, the project has evolved into a working offline-first desktop application capable of processing tender documents, extracting information, indexing evidence, providing AI-assisted responses, generating bid recommendations, and exporting reports.
Recent improvements include:
Real local Qwen2.5 inference running through llama.cpp
Offline AI operation with no dependency on cloud APIs for inference
Evidence-based tender analysis to help connect AI responses to information found in the source document
Improved dashboard state handling, ensuring that visible application information reflects the actual processing and model state
PDF export fixed and verified, alongside the existing export functionality
Tender document intelligence, including extraction of important information such as tender details, dates, requirements, and procurement-related evidence
But the project is now moving into its next stage.
Building TenderLM
I am currently developing TenderLM, a separate project focused on understanding the fundamentals of building and evaluating a language model specifically for tender and procurement intelligence.
TenderLM is not simply being added as a replacement for everything that already works in TenderGPT Edge.
The goal is to gradually explore how a purpose-built language model can eventually integrate with TenderGPT Edge and contribute to tasks such as:
Understanding tender and procurement language Analysing tender requirements Working with retrieved document evidence Supporting structured reasoning over tender information Exploring specialised local AI models for procurement intelligence
The TenderLM work has already involved building and experimenting with components including tokenization, vocabulary construction, multilingual benchmark data, model architecture, training, and evaluation.
This journey has reinforced an important lesson:
Building useful AI is not only about using the biggest available model. It is also about understanding the data, architecture, evaluation, limitations, and the environment in which the system must operate.
TenderGPT Edge remains focused on being a practical, offline-first tender intelligence application.
TenderLM represents the next research and development step: exploring whether a more specialised model can eventually become part of that system.
What's next?
The next phase is to continue developing TenderLM while carefully exploring how it can integrate with TenderGPT Edge without breaking the working evidence, document-processing, and local AI capabilities already in place.
The long-term vision is an offline-first tender intelligence system where:
Tender Documents → Extraction → Evidence → Retrieval → TenderLM / Local LLM → Explainable Tender Intelligence
Still building. Still testing. Still learning. And most importantly:
No fake results. No pretending that something works when it doesn't.
The journey from TenderGPT Edge to TenderLM has only just begun.
Log in or sign up for Devpost to join the conversation.